Evolution of robust high speed optical-flow-based landing for autonomous MAVs

Abstract

Neurocontrollers for high speed optical-flow-based landing were automatically developed with better results than the user-defined baseline.• The optimized behavior was seamlessly transferred to the real world with the aid of a closed loop control system.• Preprocessing the sensory input allowed the behavior to be tested with both a CMOS camera and Dynamic Vision Sensor with no noticeable changes to the performance.• No significant performance differences were observed despite clear differences in the neurocontroller input. A B S T R A C TAutomatic optimization of robotic behavior has been the long-standing goal of Evolutionary Robotics. Allowing the problem at hand to be solved by automation often leads to novel approaches and new insights. A common problem encountered with this approach is that when this optimization occurs in a simulated environment, the optimized policies are subject to the reality gap when implemented in the real world. This often results in sub-optimal behavior, if it works at all. This paper investigates the automatic optimization of neurocontrollers to perform quick but safe landing maneuvers for a quadrotor micro air vehicle using the divergence of the optical flow field of a downward looking camera. The optimized policies showed that a piece-wise linear control scheme is more effective than the simple linear scheme commonly used, something not yet considered by human designers. Additionally, we show the utility in using abstraction on the input and output of the controller as a tool to improve the robustness of the optimized policies to the reality gap by testing our policies optimized in simulation on real world vehicles. We tested the neurocontrollers using two different methods to generate and process the visual input, one using a conventional CMOS camera and one a dynamic vision sensor, both of which perform significantly differently than the simulated sensor. The use of the abstracted input resulted in near seamless transfer to the real world with the controllers showing high robustness to a clear reality gap.

Evolved neuromorphic radar-based altitude controller for an autonomous open-source blimp

“…The mutated offspring is then evaluated in a modelbased simulation environment (see Section III-D), where a source of random Gaussian noise is added to the radar signal. Since this randomization stimulates the persistence of controllers that are independent of such disturbances, it helps minimizing the reality gap [38]. During the evaluation, a set of 10 different reference altitudes h ref ∈ [0, 3] is provided along a total simulated duration of T = 15 seconds each.…”

Evolved Neuromorphic Control for High Speed Divergence-based Landings of MAVs

“…Regarding evolutionary robotics, [16] demonstrates realworld optical flow control of a landing MAV, where the ANN controller was evolved offline. A shallow network was sufficient to perform continuous control, with only the weights being evolved.…”

“…The vertical simulation environment in which individuals are evaluated makes use of domain randomization and artificial noise to improve transferability to the real world. The available observations are the divergence D and its temporal derivative ∆ D. Similarly to [16], the simulated MAV is considered as a unit mass under the influence of gravity, and control happens in one dimension with the SNN controller selecting a thrust setpoint T sp . This leads to the following dynamics model:…”

“…Landings start from an initial altitude of roughly 4 m. Horizontal guidance and ground-truth measurements are provided by a motion capture system. Similar to [16], [37], divergence is estimated as size divergence using the Bebop's downwardlooking CMOS camera, and a FAST corner detector [38] in combination with a pyramidal Lucas-Kanade feature tracker [39]. To limit computational expense, N D is capped at 100 points.…”

“…Artificial neural networks (ANNs), on the other hand, have been employed successfully for real-world, vision-based continuous control. For instance, [16] used an evolutionary algorithm to optimize ANNs for performing divergencebased landings of MAVs. Solving the same control problem, but replacing the ANNs with SNNs, this work aims to demonstrate that SNNs can likewise be taught to perform such a task using evolutionary learning, or neuroevolution [17].…”